The WhatsApp Bot That Grades Itself Every Night and Gets Better
Every conversation your AI bot has is scored 0–100 against your goal. Failures get specific issue lists. Prompt Improve turns those issues into patch cards you accept or reject. The bot rewrites itself — with your approval.
Every chatbot starts good. Most stay the same.
Amit Kapoor, a solar EPC owner in Jaipur, set up a WhatsApp chatbot to handle enquiries. It worked reasonably well at first. But six months later, the bot was still giving the same answers it gave on day one — even though his products had changed, his pricing had changed, and his customers were asking different questions.
He had no way to know which conversations the bot was handling well and which were frustrating leads into silence. He'd have to read every chat himself — which defeated the purpose of having a bot.
The self-improving WhatsApp bot solves this. It reads its own conversations every night, scores them, finds the patterns in the bad ones, and surfaces exactly what needs to change — as specific, editable patch cards, not vague feedback.
How nightly self-grading works
Every night, the system reviews all the bot's conversations from that day. No manual review required.
0–100 conversation score
Every conversation the bot has gets a numeric score based on how well it achieved the goal you set. A conversation where the bot confirmed a demo booking scores 95+. One where the lead asked about pricing and the bot gave a vague answer scores lower.
PASS / FAIL verdict + top issues
Each low-scoring conversation gets a FAIL verdict and a list of specific issues — e.g. 'bot missed the opportunity to ask for the lead's timeline', 'bot responded to an off-topic question instead of redirecting'. Not vague feedback; actionable specifics.
Prompt Improve patch cards
The system aggregates patterns across all FAIL conversations and turns them into patch cards — proposed edits to the bot's system prompt. Each card shows the current text and the proposed replacement. Accept to apply, reject to skip.
Accept / reject + revert
You stay in control. Accept a patch card and the bot's prompt updates immediately. If the next day's conversations suggest the patch made things worse, revert it in one click. The previous version is always saved.
Specific fixes, not vague advice
Most AI bots tell you "performance could be improved". Prompt Improve tells you exactly what to change — and lets you apply the change with one click.
Pattern analysis across failures
If 12 out of 30 conversations this week failed because the bot gave a price without first qualifying the lead's budget, that pattern shows up as one issue — not 12 individual chat reviews.
Current vs Updated diff
Each patch card shows the exact text that will change — highlighted diff between the current prompt paragraph and the proposed replacement. No guessing what's changing.
Accept to apply, reject to skip
You decide which patches to accept. Accept the ones that look right. Reject the ones you disagree with. The bot never changes without your explicit approval.
Revert if it makes things worse
Every accepted patch is saved with its before/after state. If the next day's grading shows things got worse after a patch, revert it instantly — the previous prompt version is always preserved.
What the grading looks like in practice
Real examples of how 3Sigma's AI bot evaluates a conversation against your goal.
Lead asked about pricing, bot answered from KB, redirected to demo booking, lead confirmed a slot.
Lead asked about competitors. Bot compared features but never redirected to booking a demo.
- → Bot spent 4 turns on competitor comparison without redirecting
- → No attempt to offer a demo after answering competitor question
- → Conversation ended without a next step
Lead interested but bot didn't guide them to fill the web form link shared in message 3.
- → Bot didn't follow up when lead ignored the form link
- → No reminder sent in the next message
Starting point: Prompt Builder
Before the bot can improve itself, it needs a starting prompt. 3Sigma's Prompt Builder writes the initial system prompt for you — a two-question wizard that asks about your business and what you want the bot to achieve, then drafts a full system prompt from those answers.
You can refine the generated prompt by chatting with the AI, or edit it manually. Once you go live, the self-grading cycle begins — and the bot iterates from there.
Answer 2 questions
What does your business do? What goal do you want the bot to achieve?
AI drafts your prompt
A complete system prompt generated from your answers — ready to go live or tweak.
Self-improve from day 1
Once live, every conversation is graded. Prompt Improve starts improving on the first night.
The full AI bot — self-improvement is one layer
Self-grading and Prompt Improve are built on top of 3Sigma's complete WhatsApp AI bot stack:
Knowledge base (PDF/DOCX/CSV)
Upload up to 10 files (25 MB each) as the bot's memory. It answers from your product catalogue, FAQs, or price list — not from the internet.
Test Chat before going live
Ask the bot any question your customers might ask before you switch it on. Fix gaps in the KB or the prompt before a real lead sees a wrong answer.
Gemini 2.5 Flash model
Powered by Google's Gemini 2.5 Flash — fast, capable, cost-effective. Model picker available if you want to switch to Flash-Lite or 2.5 Pro.
AI auto follow-up
The bot doesn't just reply — it follows up on its own with goal-driven messages, business-hours scheduling and three pace presets.
AI Actions (opt-in)
Bot can add a CRM note, change lead status, create a task or move the lead stage mid-conversation — with your permission for each action type.
Credit wallet
Prepaid credits: 1 credit = 1 AI message. Buy packs or top up as needed. Full activity ledger so you know where credits went.
Common questions
What does 'self-improving WhatsApp bot' mean?
Every night, 3Sigma's AI bot reviews all the conversations it had that day, grades each one from 0 to 100 against your goal (e.g. 'book a demo', 'collect documents'), and identifies what went wrong in the conversations that scored low. It then surfaces those issues as Prompt Improve patch cards — specific, targeted rewrites of the bot's system prompt. You accept or reject each one. The bot literally rewrites itself based on real conversations.
How does the nightly grading work?
After each WhatsApp conversation, the AI evaluates whether the bot achieved its goal. Each conversation gets a score from 0 to 100, a PASS or FAIL verdict, and a list of the top issues that caused it to fail (e.g. 'bot didn't ask for the budget', 'bot didn't redirect when the lead went off-topic'). The full grading report is available inside 3Sigma the next morning.
What is a Prompt Improve patch card?
A patch card is a specific proposed change to the bot's system prompt — shown as a side-by-side diff of the current text vs the proposed update. You see exactly what will change before accepting. If a patch makes things worse, you can revert it. You decide which improvements to apply, so the bot never changes without your approval.
Do I need to write prompts myself to use this?
No. 3Sigma has a Prompt Builder ('Build with AI') that generates your bot's initial system prompt from a two-question wizard about your business and goal. The self-improvement layer then iterates on that starting prompt over time. You can also write or edit prompts manually if you prefer full control.
Is this feature available on all 3Sigma accounts?
The Gemini AI bot, including the self-grading and Prompt Improve system, is available on eligible 3Sigma accounts (org-gated via isAIEnabled). Contact the 3Sigma team to check eligibility and activate the AI bot for your account.
What if the grading is wrong about a conversation?
The grading is based on whether the conversation met the goal you defined. If a conversation is graded as FAIL but you think it was actually fine, you can ignore or reject the patch card it generates. All improvements are opt-in — the bot never changes without you accepting the change.
How is this different from other WhatsApp chatbots in India?
Most WhatsApp chatbots — keyword bots, rule-based flows, even many AI bots — stay exactly the same after you set them up. You have to notice they're underperforming and rewrite the prompt yourself. 3Sigma's AI bot reviews its own conversations every night, identifies what's not working, and proposes specific fixes. No other WhatsApp chatbot platform marketed to Indian SMBs does this.
Explore the full AI bot stack
A WhatsApp bot that gets better every day — on its own
Set up the bot once. It grades itself every night, proposes specific improvements and waits for your approval before making any changes. Available on eligible 3Sigma CRM accounts.